Reinforcement Learning in the Control of a Simulated Life Support System

Abstract

To make extended space missions, such as missions to Mars, a reality, an advanced life support system (ALS) must be developed that is able to utilize resources to their fullest capabilities [2]. In order to make such a system a reality, a robust control system must be developed that is able to cope with the complexity of an ALS. This work applies reinforcement learning (RL), a machine learning technique, to the task of controlling the water recovery system of a simulated ALS. The RL agent learns an effective control strategy that extends the mission length to the point that lack of water is no longer the cause of mission termination.